We desperately need a working robust speech recognition system including sufficient capabilities to easily perform training on a certain person.

We tried it with Sphinx  http://cmusphinx.sourceforge.net/ last year but due to the tight schedule and some miscalculations we screwed up big time. One main Reason was that we weren't able to and in fact didn't train the whole thing in advance. Furthermore, even if we had trained it, the very noisy environment is quite a challenge itself.

As a first step, a speaker dependent system that works robustly is a MUST!
As an enhancement, speaker independent recognition would definitively be a big PLUS.